GANs

Question 1

GAN

Generative Adversarial Network

Answer 1

A generative model that yields high-fidelity generations with no blurring. Comprised of two fundemental components:

• Generator: uses random noise as input to generate samples similiar to those in the training set.
• Discriminator: has to discern between real photos taken from the training set and generated samples.

The discriminator learns the domain distribution just as the generator learns to mimic it. Any artifact that the generator might introduce is liable to tip off the discriminator, giving the generator a strong motive for high-fidelity and low noise.

Question 2

GAN

Generative Adversarial Network

Answer 2

A generative model that yields high-fidelity generations with no blurring. Comprised of two fundemental components:

• Generator: uses random noise as input to generate samples similiar to those in the training set.
• Discriminator: has to discern between real photos taken from the training set and generated samples.

The discriminator learns to appreciate the subtleties of the domain distribution just as the generator learns to mimic them. Any artifact that the generator might introduce is liable to tip off the discriminator, giving the generator a strong motive for high-fidelity and low noise.

Question 3

Loss Function of GAN

Answer 3

Adversarial Loss: D wants to maximize its discerment rate while G wants to minimize it. The weights are trained so as to acheive the best discernment and the best generation.
This is an unstable function. One solution to its instability is to train G and D seperately a few iterations at a time.

Also referred to as min-max loss